人民珠江2025，Vol.46Issue(5)：51-62,12.DOI:10.3969/j.issn.1001-9235.2025.05.006
耦合物理模型与深度学习的流量预测方法
Flow Prediction Method Combining Physical Model and Deep Learning:A Case Study of Gaodao Station along Lianjiang River
摘要
Abstract
This study took the"22·6"flood event at the Gaodao Station along the Lianjiang River in the middle and upper reaches of the Beijiang River in Guangdong Province as an example to explore the flow prediction method combining physical models with deep learning,aiming to improve the accuracy of hydrological predictions under extreme weather conditions.The study adopted a combination of the hydrologic engineering center-hydrologic modeling system(HEC-HMS)distributed hydrological model and the long short-term memory(LSTM)network to construct three types of coupled models,namely the HEC-LSTM model based on error correction,the HECo1-LSTM model based on single-station flow,and the HECo2-LSTM model based on multi-sub-basin output.Through prediction experiments with forecast periods of three hours,six hours,and 12 hours,the performance of each coupled model and the single hydrological model in runoff forecasting and extreme flood events was compared.The results show that the HEC-HMS model has limitations when the flow fluctuates greatly;the error correction-based HEC-LSTM model significantly improves the prediction accuracy in the short and medium term,with the root mean square error(RMSE)reduced by approximately 46%in the training set and 25%in the validation set.The HECo1-LSTM and HECo2-LSTM models perform outstandingly in long-term forecast periods,with the HECo2-LSTM model reducing the RMSE by 58%in the training set and 33%in the validation set and maintaining a high prediction accuracy(Nash-Sutcliffe model efficiency coefficient of 0.91)even in the 12-hour forecast period.This study provides a new coupling method for hydrological simulation and prediction in river basins,which is expected to significantly improve the accuracy and reliability of hydrological forecasts under extreme weather conditions.关键词
深度学习/分布式水文模型/流量预报/HEC-HMS/LSTM/连江/高道站Key words
deep learning/distributed hydrological model/flow prediction/HEC-HMS/LSTM/Lianjiang River/Gaodao Station分类
水利科学引用本文复制引用
黄泽禧,孙伟,陈新林,容泽荣,罗小康,王先伟..耦合物理模型与深度学习的流量预测方法[J].人民珠江,2025,46(5):51-62,12.基金项目
珠江人才计划青年拔尖人才项目(2019QN01G106) （2019QN01G106）
国家自然科学基金(51961125206) （51961125206）
广东省自然科学基金(2025A1515010981) （2025A1515010981）
